Field notes from the AI trenches.
How is performance between local Qwen on Apple Silicon vs. a 3090? The first layer that gets offloaded to RAM results in a fixed-cost memory copy back-and-forth to RAM. The cliff from 0->1 layers in RAM is expected. Dashed blue and orange lines are the performance on my Mac Studio (M3 Ultra). What does surprise me a bit? The stability of the results. Each point is 6 runs. The latency and…
The Gang of Four book and Working Effectively With Legacy Code are more valuable after agents than before. Legacy Code because of the proliferation of massive codebases of questionable quality. Before this, it was rare to have codebases like this. That's why Facebook used to recommend this to every new hire. And Design Patterns ? If you can identify where there's a clean seam, where there's a…
before you ask I wrote this; not a bot. Watch: I'll put in a typo to provolone it. wtf is this Where do local models (or any models?) break with injected noise for real coding workflows? We will take Qwen 9B 3.5 Q3_K_M . Throw some tasks at it and inject really long English-language novels mid-task. This is a small model. It's heavily quantized. We're going to heavily quantize its KV cache too. We…
AI disclaimer. The writing on this page is mine. The essay below, "Invalidating a business idea...." was originally written March 17, 2026. Opus gave me a few rounds of feedback on it. All the text emitted was from my keyboard though. I tried to build a company called Curtilage. It was intended to be offline AI for legal. As I reflect, I'm most impressed by the generosity of folks I spoke to. I…
This is the first blog post on my 2026 blog. Testing the ability to add screenshots with low friction: Excellent. My name is Robert Karl, I'm a software engineer and serial founder in Seattle. I studied CS at Caltech. I used to work at Facebook, Xnor.ai, and AWS. The intent with this blog is to talk about AI and building products and startups with it. Egad! Not another AI blog.
hello world Testing links between posts. This is generated with a minimal SSG, written in Obsidian, and allows easy copy paste of screenshots (critical for me!) It took a few prompts and maybe half an hour to set up. We're serving on robertkarl.net which is a Digital Ocean droplet, with nginx. Let's test our other blog features: def a(): """ Blocks of code. this one is python. Isn't whitespace…